STAGE-Claw: leaderboard
Metric: Average task score (0-100) from the verification programs, 40 automatically built and human-audited personal-agent tasks run in real operating-system environments (calendar, email, reminders, files); the final system state is checked by task-specific verification programs; OpenClaw agent framework, reasoning disabled where the model allows it, temperature 0 where configurable, first valid run per task; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Opus 4.7 | 77.1 |
| 2 | Claude Sonnet 4.6 | 69.43 |
| 3 | GPT-5.5 (Non-reasoning) | 69.19 |
| 4 | GLM-5 (Non-reasoning) | 67.4 |
| 5 | Gemini 3.1 Pro (Preview) | 65.5 |
| 6 | GPT-5.4 (Non-reasoning) | 65.37 |
| 7 | Seed 2.0 Pro (Non-reasoning) | 61.07 |
| 8 | DeepSeek V4 Pro (Non-reasoning) | 59.78 |
| 9 | Qwen 3.5 Plus (Non-reasoning) | 59.74 |
| 10 | Kimi K2.6 (Non-reasoning) | 59.05 |
| 11 | MiniMax-M2.7 | 47.53 |
Interactive version: theaggregate.ai/benchmark?slug=stage-claw · How It Works · Data refreshed daily, snapshot 2026-09-29.